Benchmarking Parallel File System Sensitiveness to I/O patterns
Konstantinos Chasapis, Jean-Yves Vet, Jean-Thomas Acquaviva · 2019
With the raise of data analytic and more generally data intensive computing, the traditional benchmarking of parallel file systems is stretched to its limits. While often promoted on their peak performance, most have difficulties to cope with the complex I/O patterns often observed in modern applications. This results in limited insights if not misleading information for end-users. The community has reacted by developing more sophisticated measurement tools and methodology, the most preeminent effort at the moment is the emerging IO500 benchmark. In this paper, we investigate the impact of the I/O access patterns on parallel file systems. Starting with the IO500 benchmark, the measurements demonstrate how specific patterns can be handled safely by some file systems, while at the opposite impact dramatically performance on others.